Deep learning for tear film stability assessment and breakup pattern classification in dry eye diagnosis

This report describes the development of an artificial intelligence system for automated assessment of tear film stability and the classification of tear breakup patterns (BUPs), supporting precise dry eye (DE) diagnosis and personalised treatment under the framework of tear film-oriented diagnosis.

A dataset of 143 eyes (98 with dry eye, 45 healthy) was analysed. Participants were evaluated using the Ocular Surface Disease Index, Keratograph 5M assessments (tear meniscus height, non-invasive tear breakup time, lipid layer thickness grade and meiboscore) and the Schirmer I test. Tear film breakup with fluorescein was video recorded and processed via a mask region-based convolutional neural network for tear film segmentation and a temporal segment network for temporal feature extraction. The proposed system performed automated evaluation of DE and identified four representative BUPs, including line break, dimple break, spot break and random break. Model performance was assessed using key metrics, including accuracy, precision, recall and F1 score.

The diagnostic model achieved 91.8% accuracy in detecting tear film breakup and 98% accuracy in identifying DE. The classification model reached 88% accuracy, with high precision and recall for line and spot break. Gradient-weighted class activation mapping indicated effective capture of morphological features. Significant differences in ocular surface parameters were observed between DE and healthy eyes, as well as among BUP groups, confirming the system's clinical relevance.

This artificial intelligence system offers an objective, non-invasive approach to supporting DE diagnosis and classifying BUPs, reducing interobserver variability and extensive training. By accurately detecting tear film instability and deficiency patterns, it supports targeted tear film-oriented therapy and advances personalised treatment strategies.